Abstract: A machine vision inspection system for realtime on-line inspection was developed, which aimed at the detection of impurities on orange petals before canned, such as hairs and fibriform objects. The hardware system was mainly made up of six CCD cameras and two computers. The software system was developed to separate the impurities from the orange sections by the color and shape feature. Some algorithms including region growing, area-reconstruction operator were used. Area-reconstruction was a new connected operator, which could preserve contours when simplifying the image. The system was able to inspect the orange petals at a rate of 30 petals/second and the inaccurate classification rate was only about 2%. The experimental results show clearly that the machine vision inspection system reaches two main objectives: accurate feature extraction and on-line treatment capabilities and can be adopted for further commercial utilization.